Abstract
Purpose:
The differential diagnosis of anterior uveitis is often complex, as it is based on the presence of different signs, symptoms, epidemiologic characteristics and ancillary tests results. The purpose of this study is to measure the diagnostic ability of a Bayesian Network-based algorithm for the etiologic diagnosis of 11 common causes of anterior uveitis (idiopathic, ankylosing spondylitis, psoriasic arthritis, reactive arthritis, inflammatory bowel diseases, sarcoidosis, tuberculosis, Behçet, Posner-Schlossman syndrome, juvenile idiopathic arthritis -JIA- and Fuchs heterochromic cyclitis).
Methods:
A retrospective validation of a diagnostic test was performed. A sample of 200 patients was randomly selected among all the patients that received a diagnosis of anterior uveitis during 2013 in a tertiary-care, hospital in London. Epidemiologic data, clinical signs and symptoms and results of ancillary tests were obtained from the patients’ clinical notes. The data were entered in the algorithm, and diagnostic results were compared to the senior clinician diagnosis (gold standard). In order to improve the estimation of the most uncommon diagnosis, the sample was enriched with 3 cases of JIA, 3 psoriasic arthritis and 4 Behçet. Robustness analysis was performed using a second algorithm with equalized pre-test probabilities for the 11 etiologies.
Results:
In 134 of 210 patients (63.8%) the most probable etiology by the algorithm matched the senior clinician diagnosis. In 169 of 210 patients (80.5%) the clinician diagnosis matched the first or second most probable results by the algorithm. Taking into account only the most probable diagnosis by the algorithm, sensitivities for each etiology ranged from 100% (7 of 7 patients with reactive arthritis and 5 of 5 with Behçet correctly classified) to 51.2% (43 of 84 patients with idiopathic anterior uveitis). Specificities ranged from 88.8% for sarcoidosis to 99.5% for Posner. A second model with equal pre-test probabilities for all 11 etiologies correctly classified 86 of 210 (41.0%) for the most probable etiology, and 108 of 210 (51.4%) for first or second most probable etiologies.
Conclusions:
The Bayesian network developed may help clinicians with the differential diagnosis of anterior uveitis. Pre-test probabilities need to be adjusted to the local population. Sensitivities and specificities of the algorithm differ among different etiologies.